lib/librte_eal/linuxapp/eal/eal.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   Copyright(c) 2012-2014 6WIND S.A.
   6  *   All rights reserved.
   7  *
   8  *   Redistribution and use in source and binary forms, with or without
   9  *   modification, are permitted provided that the following conditions
  10  *   are met:
  11  *
  12  *     * Redistributions of source code must retain the above copyright
  13  *       notice, this list of conditions and the following disclaimer.
  14  *     * Redistributions in binary form must reproduce the above copyright
  15  *       notice, this list of conditions and the following disclaimer in
  16  *       the documentation and/or other materials provided with the
  17  *       distribution.
  18  *     * Neither the name of Intel Corporation nor the names of its
  19  *       contributors may be used to endorse or promote products derived
  20  *       from this software without specific prior written permission.
  21  *
  22  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  25  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  26  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  27  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  28  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  29  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  30  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  32  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  33  */
  34
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <stdint.h>
  38 #include <string.h>
  39 #include <stdarg.h>
  40 #include <unistd.h>
  41 #include <pthread.h>
  42 #include <syslog.h>
  43 #include <getopt.h>
  44 #include <sys/file.h>
  45 #include <fcntl.h>
  46 #include <dlfcn.h>
  47 #include <stddef.h>
  48 #include <errno.h>
  49 #include <limits.h>
  50 #include <errno.h>
  51 #include <sys/mman.h>
  52 #include <sys/queue.h>
  53 #include <sys/io.h>
  54
  55 #include <rte_common.h>
  56 #include <rte_debug.h>
  57 #include <rte_memory.h>
  58 #include <rte_memzone.h>
  59 #include <rte_launch.h>
  60 #include <rte_tailq.h>
  61 #include <rte_eal.h>
  62 #include <rte_eal_memconfig.h>
  63 #include <rte_per_lcore.h>
  64 #include <rte_lcore.h>
  65 #include <rte_log.h>
  66 #include <rte_random.h>
  67 #include <rte_cycles.h>
  68 #include <rte_string_fns.h>
  69 #include <rte_cpuflags.h>
  70 #include <rte_interrupts.h>
  71 #include <rte_pci.h>
  72 #include <rte_devargs.h>
  73 #include <rte_common.h>
  74 #include <rte_version.h>
  75 #include <rte_atomic.h>
  76 #include <malloc_heap.h>
  77 #include <rte_eth_ring.h>
  78
  79 #include "eal_private.h"
  80 #include "eal_thread.h"
  81 #include "eal_internal_cfg.h"
  82 #include "eal_filesystem.h"
  83 #include "eal_hugepages.h"
  84
  85 #define OPT_HUGE_DIR    "huge-dir"
  86 #define OPT_PROC_TYPE   "proc-type"
  87 #define OPT_NO_SHCONF   "no-shconf"
  88 #define OPT_NO_HPET     "no-hpet"
  89 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  90 #define OPT_NO_PCI      "no-pci"
  91 #define OPT_NO_HUGE     "no-huge"
  92 #define OPT_FILE_PREFIX "file-prefix"
  93 #define OPT_SOCKET_MEM  "socket-mem"
  94 #define OPT_USE_DEVICE  "use-device"
  95 #define OPT_PCI_WHITELIST "pci-whitelist"
  96 #define OPT_PCI_BLACKLIST "pci-blacklist"
  97 #define OPT_VDEV        "vdev"
  98 #define OPT_SYSLOG      "syslog"
  99 #define OPT_BASE_VIRTADDR   "base-virtaddr"
 100 #define OPT_XEN_DOM0    "xen-dom0"
 101 #define OPT_CREATE_UIO_DEV "create-uio-dev"
 102
 103 #define RTE_EAL_BLACKLIST_SIZE  0x100
 104
 105 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
 106
 107 #define SOCKET_MEM_STRLEN (RTE_MAX_NUMA_NODES * 10)
 108
 109 #define HIGHEST_RPL 3
 110
 111 #define BITS_PER_HEX 4
 112
 113 /* Allow the application to print its usage message too if set */
 114 static rte_usage_hook_t rte_application_usage_hook = NULL;
 115
 116 TAILQ_HEAD(shared_driver_list, shared_driver);
 117
 118 /* Definition for shared object drivers. */
 119 struct shared_driver {
 120         TAILQ_ENTRY(shared_driver) next;
 121
 122         char    name[PATH_MAX];
 123         void*   lib_handle;
 124 };
 125
 126 /* List of external loadable drivers */
 127 static struct shared_driver_list solib_list =
 128 TAILQ_HEAD_INITIALIZER(solib_list);
 129
 130 /* early configuration structure, when memory config is not mmapped */
 131 static struct rte_mem_config early_mem_config;
 132
 133 /* define fd variable here, because file needs to be kept open for the
 134  * duration of the program, as we hold a write lock on it in the primary proc */
 135 static int mem_cfg_fd = -1;
 136
 137 static struct flock wr_lock = {
 138                 .l_type = F_WRLCK,
 139                 .l_whence = SEEK_SET,
 140                 .l_start = offsetof(struct rte_mem_config, memseg),
 141                 .l_len = sizeof(early_mem_config.memseg),
 142 };
 143
 144 /* Address of global and public configuration */
 145 static struct rte_config rte_config = {
 146                 .mem_config = &early_mem_config,
 147 };
 148
 149 /* internal configuration (per-core) */
 150 struct lcore_config lcore_config[RTE_MAX_LCORE];
 151
 152 /* internal configuration */
 153 struct internal_config internal_config;
 154
 155 /* used by rte_rdtsc() */
 156 int rte_cycles_vmware_tsc_map;
 157
 158 /* Return a pointer to the configuration structure */
 159 struct rte_config *
 160 rte_eal_get_configuration(void)
 161 {
 162         return &rte_config;
 163 }
 164
 165 /* parse a sysfs (or other) file containing one integer value */
 166 int
 167 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 168 {
 169         FILE *f;
 170         char buf[BUFSIZ];
 171         char *end = NULL;
 172
 173         if ((f = fopen(filename, "r")) == NULL) {
 174                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 175                         __func__, filename);
 176                 return -1;
 177         }
 178
 179         if (fgets(buf, sizeof(buf), f) == NULL) {
 180                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 181                         __func__, filename);
 182                 fclose(f);
 183                 return -1;
 184         }
 185         *val = strtoul(buf, &end, 0);
 186         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 187                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 188                                 __func__, filename);
 189                 fclose(f);
 190                 return -1;
 191         }
 192         fclose(f);
 193         return 0;
 194 }
 195
 196
 197 /* create memory configuration in shared/mmap memory. Take out
 198  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 199  * This means we never close the file while running (auto-close on exit).
 200  * We also don't lock the whole file, so that in future we can use read-locks
 201  * on other parts, e.g. memzones, to detect if there are running secondary
 202  * processes. */
 203 static void
 204 rte_eal_config_create(void)
 205 {
 206         void *rte_mem_cfg_addr;
 207         int retval;
 208
 209         const char *pathname = eal_runtime_config_path();
 210
 211         if (internal_config.no_shconf)
 212                 return;
 213
 214         if (mem_cfg_fd < 0){
 215                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 216                 if (mem_cfg_fd < 0)
 217                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 218         }
 219
 220         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 221         if (retval < 0){
 222                 close(mem_cfg_fd);
 223                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 224         }
 225
 226         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 227         if (retval < 0){
 228                 close(mem_cfg_fd);
 229                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 230                                 "process running?\n", pathname);
 231         }
 232
 233         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 234                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 235
 236         if (rte_mem_cfg_addr == MAP_FAILED){
 237                 rte_panic("Cannot mmap memory for rte_config\n");
 238         }
 239         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 240         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 241 }
 242
 243 /* attach to an existing shared memory config */
 244 static void
 245 rte_eal_config_attach(void)
 246 {
 247         void *rte_mem_cfg_addr;
 248         const char *pathname = eal_runtime_config_path();
 249
 250         if (internal_config.no_shconf)
 251                 return;
 252
 253         if (mem_cfg_fd < 0){
 254                 mem_cfg_fd = open(pathname, O_RDWR);
 255                 if (mem_cfg_fd < 0)
 256                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 257         }
 258
 259         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 260                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 261         close(mem_cfg_fd);
 262         if (rte_mem_cfg_addr == MAP_FAILED)
 263                 rte_panic("Cannot mmap memory for rte_config\n");
 264
 265         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 266 }
 267
 268 /* Detect if we are a primary or a secondary process */
 269 static enum rte_proc_type_t
 270 eal_proc_type_detect(void)
 271 {
 272         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 273         const char *pathname = eal_runtime_config_path();
 274
 275         /* if we can open the file but not get a write-lock we are a secondary
 276          * process. NOTE: if we get a file handle back, we keep that open
 277          * and don't close it to prevent a race condition between multiple opens */
 278         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 279                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 280                 ptype = RTE_PROC_SECONDARY;
 281
 282         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 283                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 284
 285         return ptype;
 286 }
 287
 288 /* Sets up rte_config structure with the pointer to shared memory config.*/
 289 static void
 290 rte_config_init(void)
 291 {
 292         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 293                         eal_proc_type_detect() : /* for auto, detect the type */
 294                         internal_config.process_type; /* otherwise use what's already set */
 295
 296         switch (rte_config.process_type){
 297         case RTE_PROC_PRIMARY:
 298                 rte_eal_config_create();
 299                 break;
 300         case RTE_PROC_SECONDARY:
 301                 rte_eal_config_attach();
 302                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 303                 break;
 304         case RTE_PROC_AUTO:
 305         case RTE_PROC_INVALID:
 306                 rte_panic("Invalid process type\n");
 307         }
 308 }
 309
 310 /* Unlocks hugepage directories that were locked by eal_hugepage_info_init */
 311 static void
 312 eal_hugedirs_unlock(void)
 313 {
 314         int i;
 315
 316         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 317         {
 318                 /* skip uninitialized */
 319                 if (internal_config.hugepage_info[i].lock_descriptor < 0)
 320                         continue;
 321                 /* unlock hugepage file */
 322                 flock(internal_config.hugepage_info[i].lock_descriptor, LOCK_UN);
 323                 close(internal_config.hugepage_info[i].lock_descriptor);
 324                 /* reset the field */
 325                 internal_config.hugepage_info[i].lock_descriptor = -1;
 326         }
 327 }
 328
 329 /* display usage */
 330 static void
 331 eal_usage(const char *prgname)
 332 {
 333         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 334                "[--proc-type primary|secondary|auto] \n\n"
 335                "EAL options:\n"
 336                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 337                "  -n NUM       : Number of memory channels\n"
 338                    "  -v           : Display version information on startup\n"
 339                "  -d LIB.so    : add driver (can be used multiple times)\n"
 340                "  -m MB        : memory to allocate (see also --"OPT_SOCKET_MEM")\n"
 341                "  -r NUM       : force number of memory ranks (don't detect)\n"
 342                "  --"OPT_XEN_DOM0" : support application running on Xen Domain0 "
 343                            "without hugetlbfs\n"
 344                "  --"OPT_SYSLOG"     : set syslog facility\n"
 345                "  --"OPT_SOCKET_MEM" : memory to allocate on specific \n"
 346                    "                 sockets (use comma separated values)\n"
 347                "  --"OPT_HUGE_DIR"   : directory where hugetlbfs is mounted\n"
 348                "  --"OPT_PROC_TYPE"  : type of this process\n"
 349                "  --"OPT_FILE_PREFIX": prefix for hugepage filenames\n"
 350                "  --"OPT_PCI_BLACKLIST", -b: add a PCI device in black list.\n"
 351                "               Prevent EAL from using this PCI device. The argument\n"
 352                "               format is <domain:bus:devid.func>.\n"
 353                "  --"OPT_PCI_WHITELIST", -w: add a PCI device in white list.\n"
 354                "               Only use the specified PCI devices. The argument format\n"
 355                "               is <[domain:]bus:devid.func>. This option can be present\n"
 356                "               several times (once per device).\n"
 357                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
 358                "  --"OPT_VDEV": add a virtual device.\n"
 359                "               The argument format is <driver><id>[,key=val,...]\n"
 360                "               (ex: --vdev=eth_pcap0,iface=eth2).\n"
 361                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of native RDTSC\n"
 362                "  --"OPT_BASE_VIRTADDR": specify base virtual address\n"
 363                "  --"OPT_CREATE_UIO_DEV": create /dev/uioX (usually done by hotplug)\n"
 364                "\nEAL options for DEBUG use only:\n"
 365                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 366                "  --"OPT_NO_PCI"   : disable pci\n"
 367                "  --"OPT_NO_HPET"  : disable hpet\n"
 368                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 369                "\n",
 370                prgname);
 371         /* Allow the application to print its usage message too if hook is set */
 372         if ( rte_application_usage_hook ) {
 373                 printf("===== Application Usage =====\n\n");
 374                 rte_application_usage_hook(prgname);
 375         }
 376 }
 377
 378 /* Set a per-application usage message */
 379 rte_usage_hook_t
 380 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 381 {
 382         rte_usage_hook_t        old_func;
 383
 384         /* Will be NULL on the first call to denote the last usage routine. */
 385         old_func                                        = rte_application_usage_hook;
 386         rte_application_usage_hook      = usage_func;
 387
 388         return old_func;
 389 }
 390
 391 /*
 392  * Parse the coremask given as argument (hexadecimal string) and fill
 393  * the global configuration (core role and core count) with the parsed
 394  * value.
 395  */
 396 static int xdigit2val(unsigned char c)
 397 {
 398         int val;
 399         if(isdigit(c))
 400                 val = c - '0';
 401         else if(isupper(c))
 402                 val = c - 'A' + 10;
 403         else
 404                 val = c - 'a' + 10;
 405         return val;
 406 }
 407 static int
 408 eal_parse_coremask(const char *coremask)
 409 {
 410         struct rte_config *cfg = rte_eal_get_configuration();
 411         int i, j, idx = 0 ;
 412         unsigned count = 0;
 413         char c;
 414         int val;
 415
 416         if (coremask == NULL)
 417                 return -1;
 418         /* Remove all blank characters ahead and after .
 419          * Remove 0x/0X if exists.
 420          */
 421         while (isblank(*coremask))
 422                 coremask++;
 423         if (coremask[0] == '0' && ((coremask[1] == 'x')
 424                 ||  (coremask[1] == 'X')) )
 425                 coremask += 2;
 426         i = strnlen(coremask, PATH_MAX);
 427         while ((i > 0) && isblank(coremask[i - 1]))
 428                 i--;
 429         if (i == 0)
 430                 return -1;
 431
 432         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 433                 c = coremask[i];
 434                 if (isxdigit(c) == 0) {
 435                         /* invalid characters */
 436                         return (-1);
 437                 }
 438                 val = xdigit2val(c);
 439                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 440                         if((1 << j) & val) {
 441                                 if (!lcore_config[idx].detected) {
 442                                         RTE_LOG(ERR, EAL, "lcore %u "
 443                                                 "unavailable\n", idx);
 444                                         return -1;
 445                                 }
 446                                 cfg->lcore_role[idx] = ROLE_RTE;
 447                                 if(count == 0)
 448                                         cfg->master_lcore = idx;
 449                                 count++;
 450                         } else  {
 451                                 cfg->lcore_role[idx] = ROLE_OFF;
 452                         }
 453                 }
 454         }
 455         for(; i >= 0; i--)
 456                 if(coremask[i] != '0')
 457                         return -1;
 458         for(; idx < RTE_MAX_LCORE; idx++)
 459                 cfg->lcore_role[idx] = ROLE_OFF;
 460         if(count == 0)
 461                 return -1;
 462         /* Update the count of enabled logical cores of the EAL configuration */
 463         cfg->lcore_count = count;
 464         return 0;
 465 }
 466
 467 static int
 468 eal_parse_syslog(const char *facility)
 469 {
 470         int i;
 471         static struct {
 472                 const char *name;
 473                 int value;
 474         } map[] = {
 475                 { "auth", LOG_AUTH },
 476                 { "cron", LOG_CRON },
 477                 { "daemon", LOG_DAEMON },
 478                 { "ftp", LOG_FTP },
 479                 { "kern", LOG_KERN },
 480                 { "lpr", LOG_LPR },
 481                 { "mail", LOG_MAIL },
 482                 { "news", LOG_NEWS },
 483                 { "syslog", LOG_SYSLOG },
 484                 { "user", LOG_USER },
 485                 { "uucp", LOG_UUCP },
 486                 { "local0", LOG_LOCAL0 },
 487                 { "local1", LOG_LOCAL1 },
 488                 { "local2", LOG_LOCAL2 },
 489                 { "local3", LOG_LOCAL3 },
 490                 { "local4", LOG_LOCAL4 },
 491                 { "local5", LOG_LOCAL5 },
 492                 { "local6", LOG_LOCAL6 },
 493                 { "local7", LOG_LOCAL7 },
 494                 { NULL, 0 }
 495         };
 496
 497         for (i = 0; map[i].name; i++) {
 498                 if (!strcmp(facility, map[i].name)) {
 499                         internal_config.syslog_facility = map[i].value;
 500                         return 0;
 501                 }
 502         }
 503         return -1;
 504 }
 505
 506 static int
 507 eal_parse_socket_mem(char *socket_mem)
 508 {
 509         char * arg[RTE_MAX_NUMA_NODES];
 510         char *end;
 511         int arg_num, i, len;
 512         uint64_t total_mem = 0;
 513
 514         len = strnlen(socket_mem, SOCKET_MEM_STRLEN);
 515         if (len == SOCKET_MEM_STRLEN) {
 516                 RTE_LOG(ERR, EAL, "--socket-mem is too long\n");
 517                 return -1;
 518         }
 519
 520         /* all other error cases will be caught later */
 521         if (!isdigit(socket_mem[len-1]))
 522                 return -1;
 523
 524         /* split the optarg into separate socket values */
 525         arg_num = rte_strsplit(socket_mem, len,
 526                         arg, RTE_MAX_NUMA_NODES, ',');
 527
 528         /* if split failed, or 0 arguments */
 529         if (arg_num <= 0)
 530                 return -1;
 531
 532         internal_config.force_sockets = 1;
 533
 534         /* parse each defined socket option */
 535         errno = 0;
 536         for (i = 0; i < arg_num; i++) {
 537                 end = NULL;
 538                 internal_config.socket_mem[i] = strtoull(arg[i], &end, 10);
 539
 540                 /* check for invalid input */
 541                 if ((errno != 0)  ||
 542                                 (arg[i][0] == '\0') || (end == NULL) || (*end != '\0'))
 543                         return -1;
 544                 internal_config.socket_mem[i] *= 1024ULL;
 545                 internal_config.socket_mem[i] *= 1024ULL;
 546                 total_mem += internal_config.socket_mem[i];
 547         }
 548
 549         /* check if we have a positive amount of total memory */
 550         if (total_mem == 0)
 551                 return -1;
 552
 553         return 0;
 554 }
 555
 556 static int
 557 eal_parse_base_virtaddr(const char *arg)
 558 {
 559         char *end;
 560         uint64_t addr;
 561
 562         addr = strtoull(arg, &end, 16);
 563
 564         /* check for errors */
 565         if ((errno != 0) || (arg[0] == '\0') || end == NULL || (*end != '\0'))
 566                 return -1;
 567
 568         /* make sure we don't exceed 32-bit boundary on 32-bit target */
 569 #ifndef RTE_ARCH_X86_64
 570         if (addr >= UINTPTR_MAX)
 571                 return -1;
 572 #endif
 573
 574         /* align the addr on 2M boundary */
 575         internal_config.base_virtaddr = RTE_PTR_ALIGN_CEIL((uintptr_t)addr,
 576                                                            RTE_PGSIZE_2M);
 577
 578         return 0;
 579 }
 580
 581 static inline size_t
 582 eal_get_hugepage_mem_size(void)
 583 {
 584         uint64_t size = 0;
 585         unsigned i, j;
 586
 587         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 588                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 589                 if (hpi->hugedir != NULL) {
 590                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 591                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 592                         }
 593                 }
 594         }
 595
 596         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 597 }
 598
 599 static enum rte_proc_type_t
 600 eal_parse_proc_type(const char *arg)
 601 {
 602         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 603                 return RTE_PROC_PRIMARY;
 604         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 605                 return RTE_PROC_SECONDARY;
 606         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 607                 return RTE_PROC_AUTO;
 608
 609         return RTE_PROC_INVALID;
 610 }
 611
 612 /* Parse the argument given in the command line of the application */
 613 static int
 614 eal_parse_args(int argc, char **argv)
 615 {
 616         int opt, ret, i;
 617         char **argvopt;
 618         int option_index;
 619         int coremask_ok = 0;
 620         char *prgname = argv[0];
 621         static struct option lgopts[] = {
 622                 {OPT_NO_HUGE, 0, 0, 0},
 623                 {OPT_NO_PCI, 0, 0, 0},
 624                 {OPT_NO_HPET, 0, 0, 0},
 625                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 626                 {OPT_HUGE_DIR, 1, 0, 0},
 627                 {OPT_NO_SHCONF, 0, 0, 0},
 628                 {OPT_PROC_TYPE, 1, 0, 0},
 629                 {OPT_FILE_PREFIX, 1, 0, 0},
 630                 {OPT_SOCKET_MEM, 1, 0, 0},
 631                 {OPT_PCI_WHITELIST, 1, 0, 0},
 632                 {OPT_PCI_BLACKLIST, 1, 0, 0},
 633                 {OPT_VDEV, 1, 0, 0},
 634                 {OPT_SYSLOG, 1, NULL, 0},
 635                 {OPT_BASE_VIRTADDR, 1, 0, 0},
 636                 {OPT_XEN_DOM0, 0, 0, 0},
 637                 {OPT_CREATE_UIO_DEV, 1, NULL, 0},
 638                 {0, 0, 0, 0}
 639         };
 640         struct shared_driver *solib;
 641
 642         argvopt = argv;
 643
 644         internal_config.memory = 0;
 645         internal_config.force_nrank = 0;
 646         internal_config.force_nchannel = 0;
 647         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 648         internal_config.hugepage_dir = NULL;
 649         internal_config.force_sockets = 0;
 650         internal_config.syslog_facility = LOG_DAEMON;
 651         internal_config.xen_dom0_support = 0;
 652 #ifdef RTE_LIBEAL_USE_HPET
 653         internal_config.no_hpet = 0;
 654 #else
 655         internal_config.no_hpet = 1;
 656 #endif
 657         /* zero out the NUMA config */
 658         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 659                 internal_config.socket_mem[i] = 0;
 660
 661         /* zero out hugedir descriptors */
 662         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 663                 internal_config.hugepage_info[i].lock_descriptor = -1;
 664
 665         internal_config.vmware_tsc_map = 0;
 666         internal_config.base_virtaddr = 0;
 667
 668         while ((opt = getopt_long(argc, argvopt, "b:w:c:d:m:n:r:v",
 669                                   lgopts, &option_index)) != EOF) {
 670
 671                 switch (opt) {
 672                 /* blacklist */
 673                 case 'b':
 674                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 675                                         optarg) < 0) {
 676                                 eal_usage(prgname);
 677                                 return (-1);
 678                         }
 679                         break;
 680                 /* whitelist */
 681                 case 'w':
 682                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 683                                         optarg) < 0) {
 684                                 eal_usage(prgname);
 685                                 return -1;
 686                         }
 687                         break;
 688                 /* coremask */
 689                 case 'c':
 690                         if (eal_parse_coremask(optarg) < 0) {
 691                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 692                                 eal_usage(prgname);
 693                                 return -1;
 694                         }
 695                         coremask_ok = 1;
 696                         break;
 697                 /* force loading of external driver */
 698                 case 'd':
 699                         solib = malloc(sizeof(*solib));
 700                         if (solib == NULL) {
 701                                 RTE_LOG(ERR, EAL, "malloc(solib) failed\n");
 702                                 return -1;
 703                         }
 704                         memset(solib, 0, sizeof(*solib));
 705                         strncpy(solib->name, optarg, PATH_MAX-1);
 706                         solib->name[PATH_MAX-1] = 0;
 707                         TAILQ_INSERT_TAIL(&solib_list, solib, next);
 708                         break;
 709                 /* size of memory */
 710                 case 'm':
 711                         internal_config.memory = atoi(optarg);
 712                         internal_config.memory *= 1024ULL;
 713                         internal_config.memory *= 1024ULL;
 714                         break;
 715                 /* force number of channels */
 716                 case 'n':
 717                         internal_config.force_nchannel = atoi(optarg);
 718                         if (internal_config.force_nchannel == 0 ||
 719                             internal_config.force_nchannel > 4) {
 720                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 721                                 eal_usage(prgname);
 722                                 return -1;
 723                         }
 724                         break;
 725                 /* force number of ranks */
 726                 case 'r':
 727                         internal_config.force_nrank = atoi(optarg);
 728                         if (internal_config.force_nrank == 0 ||
 729                             internal_config.force_nrank > 16) {
 730                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 731                                 eal_usage(prgname);
 732                                 return -1;
 733                         }
 734                         break;
 735                 case 'v':
 736                         /* since message is explicitly requested by user, we
 737                          * write message at highest log level so it can always be seen
 738                          * even if info or warning messages are disabled */
 739                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 740                         break;
 741
 742                 /* long options */
 743                 case 0:
 744                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 745                                 internal_config.no_hugetlbfs = 1;
 746                         }
 747                         if (!strcmp(lgopts[option_index].name, OPT_XEN_DOM0)) {
 748                 #ifdef RTE_LIBRTE_XEN_DOM0
 749                                 internal_config.xen_dom0_support = 1;
 750                 #else
 751                                 RTE_LOG(ERR, EAL, "Can't support DPDK app "
 752                                         "running on Dom0, please configure"
 753                                         " RTE_LIBRTE_XEN_DOM0=y\n");
 754                                 return -1;
 755                 #endif
 756                         }
 757                         else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 758                                 internal_config.no_pci = 1;
 759                         }
 760                         else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 761                                 internal_config.no_hpet = 1;
 762                         }
 763                         else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 764                                 internal_config.vmware_tsc_map = 1;
 765                         }
 766                         else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 767                                 internal_config.no_shconf = 1;
 768                         }
 769                         else if (!strcmp(lgopts[option_index].name, OPT_HUGE_DIR)) {
 770                                 internal_config.hugepage_dir = optarg;
 771                         }
 772                         else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 773                                 internal_config.process_type = eal_parse_proc_type(optarg);
 774                         }
 775                         else if (!strcmp(lgopts[option_index].name, OPT_FILE_PREFIX)) {
 776                                 internal_config.hugefile_prefix = optarg;
 777                         }
 778                         else if (!strcmp(lgopts[option_index].name, OPT_SOCKET_MEM)) {
 779                                 if (eal_parse_socket_mem(optarg) < 0) {
 780                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 781                                                         OPT_SOCKET_MEM "\n");
 782                                         eal_usage(prgname);
 783                                         return -1;
 784                                 }
 785                         }
 786                         else if (!strcmp(lgopts[option_index].name, OPT_USE_DEVICE)) {
 787                                 printf("The --use-device option is deprecated, please use\n"
 788                                         "--whitelist or --vdev instead.\n");
 789                                 eal_usage(prgname);
 790                                 return -1;
 791                         }
 792                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_BLACKLIST)) {
 793                                 if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 794                                                 optarg) < 0) {
 795                                         eal_usage(prgname);
 796                                         return -1;
 797                                 }
 798                         }
 799                         else if (!strcmp(lgopts[option_index].name, OPT_PCI_WHITELIST)) {
 800                                 if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 801                                                 optarg) < 0) {
 802                                         eal_usage(prgname);
 803                                         return -1;
 804                                 }
 805                         }
 806                         else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
 807                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
 808                                                 optarg) < 0) {
 809                                         eal_usage(prgname);
 810                                         return -1;
 811                                 }
 812                         }
 813                         else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 814                                 if (eal_parse_syslog(optarg) < 0) {
 815                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 816                                                         OPT_SYSLOG "\n");
 817                                         eal_usage(prgname);
 818                                         return -1;
 819                                 }
 820                         }
 821                         else if (!strcmp(lgopts[option_index].name, OPT_BASE_VIRTADDR)) {
 822                                 if (eal_parse_base_virtaddr(optarg) < 0) {
 823                                         RTE_LOG(ERR, EAL, "invalid parameter for --"
 824                                                         OPT_BASE_VIRTADDR "\n");
 825                                         eal_usage(prgname);
 826                                         return -1;
 827                                 }
 828                         }
 829                         else if (!strcmp(lgopts[option_index].name, OPT_CREATE_UIO_DEV)) {
 830                                 internal_config.create_uio_dev = 1;
 831                         }
 832                         break;
 833
 834                 default:
 835                         eal_usage(prgname);
 836                         return -1;
 837                 }
 838         }
 839
 840         /* sanity checks */
 841         if (!coremask_ok) {
 842                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 843                 eal_usage(prgname);
 844                 return -1;
 845         }
 846         if (internal_config.process_type == RTE_PROC_AUTO){
 847                 internal_config.process_type = eal_proc_type_detect();
 848         }
 849         if (internal_config.process_type == RTE_PROC_INVALID){
 850                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 851                 eal_usage(prgname);
 852                 return -1;
 853         }
 854         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 855                         internal_config.force_nchannel == 0) {
 856                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 857                 eal_usage(prgname);
 858                 return -1;
 859         }
 860         if (index(internal_config.hugefile_prefix,'%') != NULL){
 861                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 862                 eal_usage(prgname);
 863                 return -1;
 864         }
 865         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 866                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 867                                 "at the same time\n");
 868                 eal_usage(prgname);
 869                 return -1;
 870         }
 871         /* --no-huge doesn't make sense with either -m or --socket-mem */
 872         if (internal_config.no_hugetlbfs &&
 873                         (internal_config.memory > 0 ||
 874                                         internal_config.force_sockets == 1)) {
 875                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 876                                 "together with --no-huge!\n");
 877                 eal_usage(prgname);
 878                 return -1;
 879         }
 880         /* --xen-dom0 doesn't make sense with --socket-mem */
 881         if (internal_config.xen_dom0_support && internal_config.force_sockets == 1) {
 882                 RTE_LOG(ERR, EAL, "Options --socket-mem cannot be specified "
 883                                         "together with --xen_dom0!\n");
 884                 eal_usage(prgname);
 885                 return -1;
 886         }
 887
 888         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 889                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 890                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 891                         "[-w] options cannot be used at the same time\n");
 892                 eal_usage(prgname);
 893                 return -1;
 894         }
 895
 896         if (optind >= 0)
 897                 argv[optind-1] = prgname;
 898
 899         /* if no memory amounts were requested, this will result in 0 and
 900          * will be overriden later, right after eal_hugepage_info_init() */
 901         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 902                 internal_config.memory += internal_config.socket_mem[i];
 903
 904         ret = optind-1;
 905         optind = 0; /* reset getopt lib */
 906         return ret;
 907 }
 908
 909 static void
 910 eal_check_mem_on_local_socket(void)
 911 {
 912         const struct rte_memseg *ms;
 913         int i, socket_id;
 914
 915         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 916
 917         ms = rte_eal_get_physmem_layout();
 918
 919         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 920                 if (ms[i].socket_id == socket_id &&
 921                                 ms[i].len > 0)
 922                         return;
 923
 924         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 925                         "memory on local socket!\n");
 926 }
 927
 928 static int
 929 sync_func(__attribute__((unused)) void *arg)
 930 {
 931         return 0;
 932 }
 933
 934 inline static void
 935 rte_eal_mcfg_complete(void)
 936 {
 937         /* ALL shared mem_config related INIT DONE */
 938         if (rte_config.process_type == RTE_PROC_PRIMARY)
 939                 rte_config.mem_config->magic = RTE_MAGIC;
 940 }
 941
 942 /*
 943  * Request iopl privilege for all RPL, returns 0 on success
 944  */
 945 static int
 946 rte_eal_iopl_init(void)
 947 {
 948         return iopl(HIGHEST_RPL);
 949 }
 950
 951 /* Launch threads, called at application init(). */
 952 int
 953 rte_eal_init(int argc, char **argv)
 954 {
 955         int i, fctret, ret;
 956         pthread_t thread_id;
 957         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 958         struct shared_driver *solib = NULL;
 959         const char *logid;
 960
 961         if (!rte_atomic32_test_and_set(&run_once))
 962                 return -1;
 963
 964         logid = strrchr(argv[0], '/');
 965         logid = strdup(logid ? logid + 1: argv[0]);
 966
 967         thread_id = pthread_self();
 968
 969         if (rte_eal_log_early_init() < 0)
 970                 rte_panic("Cannot init early logs\n");
 971
 972         if (rte_eal_cpu_init() < 0)
 973                 rte_panic("Cannot detect lcores\n");
 974
 975         fctret = eal_parse_args(argc, argv);
 976         if (fctret < 0)
 977                 exit(1);
 978
 979         if (internal_config.no_hugetlbfs == 0 &&
 980                         internal_config.process_type != RTE_PROC_SECONDARY &&
 981                         internal_config.xen_dom0_support == 0 &&
 982                         eal_hugepage_info_init() < 0)
 983                 rte_panic("Cannot get hugepage information\n");
 984
 985         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 986                 if (internal_config.no_hugetlbfs)
 987                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 988                 else
 989                         internal_config.memory = eal_get_hugepage_mem_size();
 990         }
 991
 992         if (internal_config.vmware_tsc_map == 1) {
 993 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 994                 rte_cycles_vmware_tsc_map = 1;
 995                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 996                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 997 #else
 998                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 999                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
1000 #endif
1001         }
1002
1003         rte_srand(rte_rdtsc());
1004
1005         rte_config_init();
1006
1007         if (rte_eal_iopl_init() == 0)
1008                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
1009
1010         if (rte_eal_pci_init() < 0)
1011                 rte_panic("Cannot init PCI\n");
1012
1013 #ifdef RTE_LIBRTE_IVSHMEM
1014         if (rte_eal_ivshmem_init() < 0)
1015                 rte_panic("Cannot init IVSHMEM\n");
1016 #endif
1017
1018         if (rte_eal_memory_init() < 0)
1019                 rte_panic("Cannot init memory\n");
1020
1021         /* the directories are locked during eal_hugepage_info_init */
1022         eal_hugedirs_unlock();
1023
1024         if (rte_eal_memzone_init() < 0)
1025                 rte_panic("Cannot init memzone\n");
1026
1027         if (rte_eal_tailqs_init() < 0)
1028                 rte_panic("Cannot init tail queues for objects\n");
1029
1030 #ifdef RTE_LIBRTE_IVSHMEM
1031         if (rte_eal_ivshmem_obj_init() < 0)
1032                 rte_panic("Cannot init IVSHMEM objects\n");
1033 #endif
1034
1035         if (rte_eal_log_init(logid, internal_config.syslog_facility) < 0)
1036                 rte_panic("Cannot init logs\n");
1037
1038         if (rte_eal_alarm_init() < 0)
1039                 rte_panic("Cannot init interrupt-handling thread\n");
1040
1041         if (rte_eal_intr_init() < 0)
1042                 rte_panic("Cannot init interrupt-handling thread\n");
1043
1044         if (rte_eal_timer_init() < 0)
1045                 rte_panic("Cannot init HPET or TSC timers\n");
1046
1047         eal_check_mem_on_local_socket();
1048
1049         rte_eal_mcfg_complete();
1050
1051         TAILQ_FOREACH(solib, &solib_list, next) {
1052                 RTE_LOG(INFO, EAL, "open shared lib %s\n", solib->name);
1053                 solib->lib_handle = dlopen(solib->name, RTLD_NOW);
1054                 if (solib->lib_handle == NULL)
1055                         RTE_LOG(WARNING, EAL, "%s\n", dlerror());
1056         }
1057
1058         eal_thread_init_master(rte_config.master_lcore);
1059
1060         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%x)\n",
1061                 rte_config.master_lcore, (int)thread_id);
1062
1063         if (rte_eal_dev_init() < 0)
1064                 rte_panic("Cannot init pmd devices\n");
1065
1066         RTE_LCORE_FOREACH_SLAVE(i) {
1067
1068                 /*
1069                  * create communication pipes between master thread
1070                  * and children
1071                  */
1072                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
1073                         rte_panic("Cannot create pipe\n");
1074                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
1075                         rte_panic("Cannot create pipe\n");
1076
1077                 lcore_config[i].state = WAIT;
1078
1079                 /* create a thread for each lcore */
1080                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
1081                                      eal_thread_loop, NULL);
1082                 if (ret != 0)
1083                         rte_panic("Cannot create thread\n");
1084         }
1085
1086         /*
1087          * Launch a dummy function on all slave lcores, so that master lcore
1088          * knows they are all ready when this function returns.
1089          */
1090         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
1091         rte_eal_mp_wait_lcore();
1092
1093         return fctret;
1094 }
1095
1096 /* get core role */
1097 enum rte_lcore_role_t
1098 rte_eal_lcore_role(unsigned lcore_id)
1099 {
1100         return (rte_config.lcore_role[lcore_id]);
1101 }
1102
1103 enum rte_proc_type_t
1104 rte_eal_process_type(void)
1105 {
1106         return (rte_config.process_type);
1107 }
1108
1109 int rte_eal_has_hugepages(void)
1110 {
1111         return ! internal_config.no_hugetlbfs;
1112 }